Voice analysis for detecting patients with Parkinson s disease using the hybridization of the best acoustic features

Transcription

1 International Journal on Electrical Engineering and Informatics - Volume 8, umber, March 206 Voice analysis for detecting patients with Parkinson s disease using the hybridization of the best acoustic features Achraf BEBA, Abdelilah JILBAB 2 and Ahmed HAMMOUCH 3,2,3 Laboratoire de Recherche en Génie Electrique, Ecole ormale Supérieure de l Enseignement Technique, Mohammed V University, Rabat, Morocco achraf.benba@um5s.net.ma Abstract: Parkinson's disease (PD) is a neurodegenerative disorder of unknown etiology. It causes, during its course, vocal impairment in approximately 90% of patients. PD patients suffer from hypokinetic dysarthria, which manifests in all aspects of speech production: respiration, phonation, articulation, nasality and prosody. To evaluate these, clinicians have adopted perceptual methods, based on acoustic cues, to distinguish different disease states. In order to improve these evaluations, we used a variety of voice samples comprising the numbers from to 0, four rhymed sentences, nine Turkish words plus the sustained vowels a, o, and u. Samples were collected from 40 people, 20 with PD. We used the method of Leave- One-Subject-Out (LOSO) validation with a K earest eighbor (k-) classifier with its different types of kernels, (i.e.; RBF, Linear, polynomial and MLP). The best result obtained was 82.5% using two diffirent voice samples; - the 4 th acoustic features along with the 7 th voice samples; 2- The 3 th and the 5 th acoustic features along with 20 th voice sample. Keywords: Voice analysis, Parkinson s disease, acoustic features, UCI Machine learning, Leave One Subject Out, Support Vector Machines.. Introduction Parkinson s disease (PD) is the second most common neurological disorder after Alzheimer's disease. It causes, during its course, a variety of symptoms. These include difficulty walking, talking, thinking or completing other simple tasks [] [2] [3]. Such neurological diseases profoundly affect the lives of patients and their families []. PD is generally seen in people over the age of 50. For most elderly people who are suffering from the disease, physical visits to the clinic for diagnosis, monitoring, and treatment are difficult and complicated [4] [5]. Parkinson s disease causes vocal impairment for approximately 90% of patients [6]. Vocal disorders do not appear abruptly. They are the result of a slow process whose early stages may be unnoticed. For this reason, the development of early diagnosis and tele-monitoring systems with accurate, reliable, and unbiased predictive models, are crucial for patients and research [] [7]. These will allow practitioners and patients to act faster and better understand the disease. There are recent studies for the detection of voice disorders with machine learning tools using acoustic measurements (features) of dysphonia. These include fundamental frequency or pitch of vocal oscillation (F0); absolute sound pressure level (indicating the relative loudness of speech); jitter which represents the cycle-to-cycle variation of fundamental frequency; shimmer which represents the extent of variation in speech amplitude from cycle to cycle; and harmonicity which represents the degree of acoustic periodicity [] [8] [9]. Studies have shown variations in all these measurements for distinguish person with PD from healthy controls [0], which shows that these acoustic parameters could be useful to evaluate speech disorders []. Voice impairments can be detected by using acoustic features extracted from people with PD. Other measurements, such as complex nonlinear aperiodicity, turbulent, aero-acoustic, and non-gaussian randomness of the sound could be useful to increase the efficacy of voice disorder diagnosis systems [8]. Received: October 29 th, 204. Accepted: February 2 nd, 206 DOI: /ijeei

2 Achraf BEBA, et al. As for disorders, Little et al. [] aimed to discriminate healthy people from people with PD by detecting dysphonia. In their study, sustained vowel a phonations were recorded from 3 subjects, of whom 23 were diagnosed with PD. They then selected ten highly uncorrelated measures, and found four that, in combination, lead to overall correct classification performance of 9.4%, using a kernel Support Vector Machine (SVM). Betul Erdogdu Sakar et al [5] analyzed multiple types of sound recordings collected from people with Parkinson s disease. The extracted features were fed into SVM and k-earest eighbor (k-) classifiers for PD diagnosis by using a leave-one-subject-out (LOSO) cross-validation scheme and summarized Leave-One-Out. To distinguish healthy subjects from PWP, most studies use SVM classification [] [5]. Success of the diagnostic system is measured with true positive (TP), true negative (T), false positive (FP) and false negative (F) rates. In this work, we used a dataset that was published in the UCI machine learning archive on June 204 and which was used by Betul Erdogdu Sakar et al [5]. In their study, multiple voice samples per subject were collected during the pronunciation of numbers from to 0, four rhymed sentences, nine words in Turkish language along with sustained vowels a, o, and u from 40 people, 20 with Parkinson s disease. In this paper we used the same database as [5]. The main idea of our work is to show the effectiveness of using, not multiple types of voice recording as [5], but each type of voice recording independently in order to improve the classification accuracy. For classification, we used k-, and to evaluate the success of the models in discriminating healthy subjects from people with Parkinson s disease, we calculated accuracy, specificity, sensitivity scores [5]. 2. Data acquisition The database of this study (Table I) was downloaded from the UCI machine learning archive, and was used in [5]. It consists of 20 Parkinsonian patients (6 female, 4 male) and 20 healthy subjects (0 female, 0 male) (Figureure ) who visited the Department of eurology in the Cerrahpasa Faculty of Medicine, Istanbul University. The test group consisted of patients suffering from Parkinson's disease for 0 to 6 years. The age of PD patients varied between 43 and 77 years (mean: 64.86, standard deviation: 8.97). The age of healthy subjects ranged between 45 and 83 years (mean: 62.55, standard deviation: 0.79). For each individual, 26 voice samples including sustained vowels, numbers, words, and short sentences were recorded. The voice samples were selected by a group of neurologists from a set of speaking exercises that aim to improve voice performance. All samples were recorded by a Trust MC-500 microphone with a frequency range between 50 Hz and 3 khz. The microphone was set to 96 khz, 30 db and placed at a distance of 0 cm from the subject, who then read or repeated the specified words or texts. In the UCI machine learning archive dataset [], there are 26 voice samples, with multiple types of voice recordings. The dataset is organized in a way that the columns represent the 26 features, and the rows represent the 26 types of voice samples for each individual, for instance the first three samples represent the sustained vowels /a/, /o/ and /u/, respectively. The samples from 4 to 3 represent numbers from to 0. The 4 th to the 7 th voice sample represent short sentences, while the rest of samples represent individual words. This makes a metrics of 040x26 (40x26 =040). 09

3 Voice analysis for detecting patients with Parkinson s disease using the hybridization Figure. Waveform of a voice sample belonging to a healthy individual (top) and a PWP (bottom). Table. Structure of the database Database (040x26) 26 Acoustic features (see Table I) which contains (F0, Jitter, Shimmer, HR) Subjects Samples Table 2. Time-Frequency-Based Features given by Pratt acoustic analysis software Groups Features umber of feature Jitter Parameters Shimmer Parameters Harmonicity Pitch Parameters Pulses Parameters Voicing Parameters Jitter (local) (%) Jitter (local, absolute) (s) 2 Jitter (rap) (%) 3 Jitter (ppq5) (%) 4 Jitter (ddp) (%) 5 Shimmer (local) (%) 6 Shimmer (local, db) (db) 7 Shimmer (apq3) (%) 8 Shimmer (apq5) (%) 9 Shimmer (apq) (%) 0 Shimmer (dda) (%) Mean autocorrelation (AC) 2 Mean HR 3 Mean HR 4 Median pitch (Hz) 5 Mean pitch (Hz) 6 Standard deviation (Hz) 7 Minimum pitch (Hz) 8 Maximum pitch (Hz) 9 umber of pulses 20 umber of periods 2 Mean period (s) 22 Standard deviation of period (s) 23 Fraction of locally unvoiced pitch frames (%) 24 umber of voice breaks 25 Degree of voice breaks (%) 26 0

4 Achraf BEBA, et al. 3. Methodology In their study, Betul Erdogdu Sakar et al [5] used a classification with Leave-One-Subject- Out (LOSO) validation scheme, in which all the 26 voice samples of one individual were left out to be used for validation as if it was an unseen individual, and the rest of the samples are used for training [5]. According to their method, if the majority of the voice samples of a test individual are classified as unhealthy, then the individual is classified as positive [5]. In their study, they presented another classification with Summarized Leave-One-Out (s-loo). The aim of using this method was to compare the success of conventional Leave-One-Subject-Out validation with an unbiased Leave-One-Out (LOO) [5]. In this method, the feature values of the 26 voice samples of each individual are summarized using central tendency and dispersion metrics such as mean, median, trimmed mean (0% and 25% removed), standard deviation, interquartile range, mean absolute deviation, and a novel form of dataset consisting of samples is formed where is the number of individuals [5]. The purpose of summarizing the voice samples of individuals is to minimize the effect of variations between different voice samples of a subject [5]. The best classification accuracy achieved in their research was 77.50% using s-loo with linear kernel of SVM and the best results of 000 runs of selecting a random voice samples from each individual was 85% with the same SVM kernel [5] which is not a steady results, because it needs 000 runs maybe less maybe more and also they used random voice samples and not the same sample for all subjects. The methodology of our method is described in the next paragraphs: A. Feature Extraction Dysarthria is the set of voice illnesses related with turbulences of muscular control of the speech organs. Dysarthria includes all malfunctions related to breathing, phonation, articulation, nasalization and prosody. These deficits can be measured and detected by analyzing various features of voice. In this study each subject was asked to read or say predetermined 26 items comprising numbers from to 0, four rhymed sentences, nine words in Turkish language plus sustained vowels a, o, and u. A total of 040 recordings each represented with by a 26 dimensional feature vector was calculated (Table II) along with a binary PD-score. A PD-score of zero indicates that the feature vector belongs to a person with PD and a score of one indicates that it belongs to a healthy subject. To extract features from voice samples, Praat acoustic analysis software [2] was used. A group of 26 linear and timefrequency based features were extracted from each voice sample.. Fundamental Frequency The fundamental frequency is the number of opening and closing cycles of the vocal folds per second. 2. Voice breaks ormal voices can easily maintain phonation when saying sustained vowel /a/. Some pathological voices have difficulty with it. This can be measured in different ways [2]. Fraction of locally unvoiced pitch frames (FLUPF) represents the fraction of pitch frames that are considered as unvoiced (the MDVP calls it DUV) [2]. umber of voice breaks is defined as the number of times distances between consecutive pulses that are longer than.25 divided by the pitch floor [2]. Degree of voice breaks is computed as the total duration of the breaks between the voiced parts of the signal, divided by the total duration of the analyzed part of the signal (the MDVP calls it DVB). [2]. 3. Jitter measurement Jitter (absolute) represents the cycle-to-cycle variation of fundamental frequency. It is computed as the average absolute difference between consecutive periods, expressed as [5] [6]:

5 Voice analysis for detecting patients with Parkinson s disease using the hybridization Jitter ( absolu) T i T i () Where T i is the extracted F 0 (fundamental frequency) period lengths and is the number of extracted F 0 periods. Jitter (relative) is defined as the average absolute difference between consecutive periods, divided by the average period. It is expressed as a percentage [5] [6]: Jitter i ( relative ) T T T i (2) Jitter (rap) represents the Relative Average Perturbation, computed as the average absolute difference between a period and the average of it and its two neighbors, divided by the average period [5] [6]. Jitter (ppq5) represents the five-point Period Perturbation Quotient, defined as the average absolute difference between a period and the average of it and its four closest neighbors, divided by the average period [5] [6]. 4. Shimmer measurement Shimmer (db) represents the variability of the peak-to-peak amplitude in decibels, computed as the average absolute base-0 logarithm of the difference between the amplitudes of consecutive periods, multiplied by 20 [5] [6]: Shimmer ( db) A 20log Ai (3) where A i is the extracted peak-to-peak amplitude data, and is the number of extracted fundamental frequency periods F 0. Shimmer (relative) is expressed as the average absolute difference between the amplitudes of consecutive periods, divided by the average amplitude, expressed as a percentage [5] [6]: Shimmer( relative ) i A A A i (4) Shimmer (apq) represents the -point Amplitude Perturbation Quotient, defined as the average absolute difference between the amplitude of a period, and the average of the amplitudes of it, and its ten closest neighbors, divided by the average amplitude [5]. 5. Harmonicity Harmonics-to-oise Ratio (HR) is expressed as the degree of acoustic periodicity [2]. B. Classification with Leave-One-Subject-Out (LOSO) 2

6 Achraf BEBA, et al. Instead of using conventional bootstrapping or LOO validation methods [3] [4] (which consist of sparing some samples of an individual in the training phase and some for the testing phase, creating an artificial overlap between the training and test sets) or s-loo (which summarizes the feature values of voice samples of each subject by using central tendency and dispersion metrics [5]), we use a LOSO validation scheme. That is, we left out all the samples of one individual to be used for validation as if it were an unseen individual, and trained a classifier on the rest of the samples. In addition, we compared two methods of using LOSO validation: first with all the 26 voice samples of each individual; then, using LOSO for each voice sample independently. Evaluation Metrics In order to measure the success of our classifiers and select the best acoustic features, we used the evaluation metrics accuracy, sensitivity and specificity. Accuracy is the ratio of correctly classified instances to total instances [5]: TP T Accuracy TP T FP F (5) Where TP is true positive (Healthy subjects who was correctly classified), T is true negative (Patients with PD who was correctly classified), FP is false positive (Patients with PD who was incorrectly classified) and F is false negative (Healthy subjects who were incorrectly classified). Accuracy represents the success of the classifier to discriminate between the two groups, Sensitivity represents the accuracy of detecting Healthy subjects and Specificity represents the accuracy of detecting the other patients with PD [5] [7] [8] [9] [20]: TP Sensitivity TP F (6) T Specifisit y T FP (7) 4. Experimental result In this study, we replicated the work in [5]. Firstly, we used LOSO method with all the 26 voice samples at the same time of each individual as can be seen in Table I. These 26 voice samples contain sustained vowels, numbers, words, and short sentences. For example; in the first step of LOSO method, we left out the 26 voice samples of the first subject to be used for validation, and we trained our classifier on the other voice samples of the other 39 subjects (all subjects except the first one). And then we decide if it is TP, T, FP or F. In the second step, we left out the 26 voice samples of the second subject to be used for validation, and we trained our classifier on the other voice samples of the other 39 subjects (all subjects except the second one) then we decide if it is TP, T, FP or F. This method is repeated for all until the 40 th subject. Then we calculated the accuracy sensitivity and specificity. The best obtained classification accuracy was % using k- with k=3 as shown in the Table III. This is somewhat lower than that what was obtained in [5] using the entire 26 voice sample at the same time. By using s-loo, they got 85% as maximum classification accuracy after 000 runs which is not a steady result, because it needs 000 runs maybe less maybe more and also they used random voice samples and not the same sample for all subjects. In order to build a real system to diagnosis the disease we should use the same recordings for all subjects and not different ones. The best obtained results using same recording was 77.50% using s-loo with linear kernel of SVM and 82.4% using s-loo with RBF kernel of SVM for the test dataset which is not use in this study. 3

7 Voice analysis for detecting patients with Parkinson s disease using the hybridization Table 3. Classification results using LOSO with all 26 voice samples k- Accuracy % Sensitivity % Specificity % Figure 2. The classification accuracies using Leave-One-Subject-Out (LOSO) validation scheme for the 20 th voice sample. The best results was achieved for the 3 th and the 5 nd voice samples (which correspond to Jitter (rap) and Jitter (ddp) successively) using LOSO with Linear k- (k = 3) Figure 3. The classification accuracies using Leave-One-Subject-Out (LOSO) validation scheme for the 7 th voice sample. The best results was achieved for the 4 th (which correspond to Jitter (ppq5)) using LOSO with Linear k- (k = 3) In our study, instead of using all the 26 voice samples at the same time, we used LOSO for every voice sample independently. That is we used LOSO for the first voice sample of each subject and we calculated the accuracy, and then for the second voice sample and so on until the 26 th voice sample. As can be seen from Figureure 2 and Figureure 3 and Table II, the best result obtained was 82.5% using two different voice samples; - the 4 th acoustic features (Jitter (ppq5)) along with the 7 th voice samples (short sentence); 2- The 3 th (Jitter (rap)) and the 5 th (Jitter (ddp)) acoustic features along with 20 th voice sample (word). These samples correspond 4

8 Achraf BEBA, et al. to sentences and words. From these results, it is clear that Jitter measurements and the running speech contain more information about the state of a subject as pathological or healthy. 5. Conclusion The characteristic disorders of Parkinson s disease are the result of a slow process whose early stages may go unnoticed. For this reason, we used a variety of voice samples per subject with multiple types of voice recordings, in order to develop method for early diagnosis and to build predictive tele-diagnosis and tele-monitoring models. The representations of acoustic signals have allowed us to examine the effect of Parkinsonism on the phonological system. However, during the various stages of processing, the results show that some voice samples are not reliable indicator of the state of a subject. Indeed, for Parkinson s patients some parameters do not confirm the presence of the disease and do not distinguish them from healthy subjects. The purpose of this study is to show the effectiveness of using each type of voice recording and each acoustic feature independently. The best result obtained was 82.5% using running speech samples along with jitter measurement by k- and LOSO validation scheme with k=3. 6. Acknowledgment The authors would like to thank Thomas R. Przybeck and Daniel Wood, United States Peace Corps Volunteers (Morocco ), and all of the participants involved in this study. 7. References: []. Little, Max A., et al. "Suitability of dysphonia measurements for telemonitoring of Parkinson's disease." Biomedical Engineering, IEEE Transactions on 56.4 (2009): [2]. Ishihara, L., and C. Brayne. "A systematic review of depression and mental illness preceding Parkinson's disease." Acta eurologica Scandinavica 3.4 (2006): [3]. Jankovic, Joseph. "Parkinson s disease: clinical features and diagnosis."journal of eurology, eurosurgery & Psychiatry 79.4 (2008): [4]. Huse, Daniel M., et al. "Burden of illness in Parkinson's disease." Movement disorders 20. (2005): [5]. Sakar, Betul Erdogdu, et al. "Collection and Analysis of a Parkinson Speech Dataset With Multiple Types of Sound Recordings." Biomedical and Health Informatics, IEEE Journal of 7.4 (203): [6]. S. B. O Sullivan and T. J. Schmitz, Parkinson disease, in Physical Rehabilitation, 5th ed. Philadelphia, PA, USA: F. A. Davis Company, 2007, pp , pp [7]. Ruggiero, C., R. Sacile, and M. Giacomini. "Home telecare." Journal of Telemedicine and Telecare 5. (999): -7. [8]. Little, Max A., et al. "Exploiting nonlinear recurrence and fractal scaling properties for voice disorder detection." BioMedical Engineering OnLine 6. (2007): 23. [9]. Rahn III, Douglas A., et al. "Phonatory impairment in Parkinson's disease: evidence from nonlinear dynamic analysis and perturbation analysis." Journal of Voice 2. (2007): [0]. Zwirner, Petra, Thomas Murry, and Gayle E. Woodson. "Phonatory function of neurologically impaired patients." Journal of communication disorders 24.4 (99): []. UCI Machine Learning Repository, Parkinson speech dataset with multiple types of sound recording data set. Available : [ archive.ics.uci.edu/ml/datasets/parkinson+speech+dataset+with++multiple+types+of+s ound+recordings] [2]. Boersma, Paul. "Praat, a system for doing phonetics by computer." Glot international 5.9/0 (2002):

9 Voice analysis for detecting patients with Parkinson s disease using the hybridization [3]. Efron, Bradley. "Bootstrap methods: another look at the jackknife." The annals of Statistics (979): -26. [4]. Reunanen, Juha. "Overfitting in making comparisons between variable selection methods." The Journal of Machine Learning Research 3 (2003): [5]. Farrús, Mireia, Javier Hernando, and Pascual Ejarque. "Jitter and shimmer measurements for speaker recognition." ITERSPEECH [6]. Shirvan, R. Arefi, and E. Tahami. "Voice analysis for detecting Parkinson's disease using genetic algorithm and K classification method." Biomedical Engineering (ICBME), 20 8th Iranian Conference of. IEEE, 20. [7]. BEBA, Achraf, Abdelilah JILBAB, and Ahmed HAMMOUCH. "Voice analysis for detecting persons with Parkinson s disease using MFCC and VQ." The 204 International Conference on Circuits, Systems and Signal Processing [8]. Benba, Achraf, Abdelilah Jilbab, and Ahmed Hammouch. "Hybridization of best acoustic cues for detecting persons with Parkinson's disease." Complex Systems (WCCS), 204 Second World Conference on. IEEE, 204. [9]. Benba, Achraf, Abdelilah Jilbab, and Ahmed Hammouch. "Voiceprint analysis using Perceptual Linear Prediction and Support Vector Machines for detecting persons with Parkinson s disease." the 3rd International Conference on Health Science and Biomedical Systems (HSBS'4), ovember , Florence, Italie. [20]. Benba, Achraf, Abdelilah Jilbab, and Ahmed Hammouch. "Voice analysis for detecting persons with Parkinson s disease using PLP and VQ." Journal of Theoretical & Applied Information Technology 70.3 (204). disorders. BEBA Achraf, received the Master degree in Electrical Engineering from Ecole ormale Supérieure de l Enseignement Technique ESET, Rabat Mohammed V University, Morocco, in 203 he is a research student of Sciences and Technology of the Engineer in Ecole ationale Supérieure d Informatique et d Analyse des Systèmes ESIAS, Research Laboratory in Electrical Engineering LRGE, Research Team in Computer and Telecommunication ERIT at ESET, Mohammed V University, Rabat, Morocco. His interests are in Signal processing for detection neurological Abdelilah JILBAB is a teacher at the Ecole ormale Supérieure de l Enseignement Technique de Rabat, Morocco; He acquired his PhD in Computer and Telecommunication from Mohammed V-Agdal University, Rabat, Morocco in February His thesis is concerned with the Filtering illegal sites on the Internet: Contribution to the type of image recognition based on the Principle of Maximum Entropy. Since 2003 he is a member of the laboratory LRIT (Unit associated with the CRST, FSR, Mohammed V University, Rabat, Morocco). Ahmed HAMMOUCH received the master degree and the PhD in Automatic, Electrical, Electronic by the Haute Alsace University of Mulhouse (France) in 993 and the PhD in Signal and Image Processing by the Mohammed V Agdal University of Rabat (Morocco) in From 993 to 203 he was professor in the Mohammed V-Souissi University in Morocco. Since 2009 he manages the Research Laboratory in Electronic Engineering. He is an author of several papers in international journals and conferences. His domains of interest include multimedia data processing and telecommunications. He is currently head of Department for Scientific and Technical Affairs in ational Center for Scientific and Technical Research in Rabat (Morocco). 6